Performance

Runtime depends more on the numerical mode and multivariate dimension than on the number of output rows alone.

Main cost drivers

Setting or input	Effect
Number of formation dates	Repeats the complete rolling calculation
Number of assets	Increases multivariate projection and transport work
Number and length of frequencies	Adds empirical measures and samples
n_projections	Approximately linear work in sliced/projected calculations
n_quantiles	Increases projected-quantile work and resolution
barycenter_size	Increases free-support optimization and transport matrices
barycenter_max_iter	Caps free-support iterations
distance="exact"	Solves discrete transport rather than projected approximations

Exact complexity also depends on the POT, NumPy, scikit-learn, and linear algebra implementations in the installed environment.

Practical workflow

1. Use validate_path_inputs to eliminate input problems without geometry.
2. Develop the pipeline with SignalConfig.projected and a modest number of projections and quantiles.
3. Benchmark a representative subset of dates and assets.
4. Increase precision settings as a documented convergence analysis.
5. Run the declared research configuration and retain its exact config JSON.

Do not tune settings only until a backtest looks attractive. Numerical convergence and economic model selection are separate questions.

Progress reporting

Pass progress_callback to integrate a logger or progress bar. The callback is invoked after each estimated or skipped formation and receives completed count, total count, date, status, and optional skip reason.

Parallelism

The public path API currently evaluates formation dates sequentially. This preserves a simple deterministic execution model and avoids imposing a process backend. If an external research system parallelizes independent point estimates, it must preserve formation-specific seeds, asset order, configuration identity, and deterministic output ordering.

Memory

The source daily panel, selected rolling window, frequency measures, and transport work arrays coexist during an estimate. Free-support exact transport can be materially heavier than sliced or projected modes. Benchmark with the largest intended asset universe and lookback rather than extrapolating only from a toy example.